Anthropic is contributing $1.5 million to the Python Software Foundation in a two-year partnership deal.

https://pyfound.blogspot.com/2025/12/anthropic-invests-in-python.html

Classic GPT 5.2 behavior (it's still a good model though)

source

GPT 5.2 Codex is now available in the Responses API.

https://platform.openai.com/docs/models/gpt-5.2-codex

OpenAI announced a new API standard for LLMs to help agentic systems, built on top of their Responses API.

https://www.openresponses.org/

New experimental Codex (by OpenAI) feature - steering. Now, if you send a message while the agent is actively doing something, the message will be sent instantly and the agent will see it mid-work, hence you can "steer" it away from doing something wrong, or provide missing details.

source

OpenAI is partnering with Cerebras (a company doing fast model inference)

https://openai.com/index/cerebras-partnership/

The capacity will come online in multiple tranches through 2028.

However, a recent tweet by Sam (and a quote-tweet by Tibo, who works on Codex) seem to indicate that it might come way sooner for Codex.

Codex 0.88.0 is out and finally has multi-agents available in /experimental

https://github.com/openai/codex/releases/tag/rust-v0.88.0

OpenAI is going to release more things related to Codex soon, including a new model that reaches 'High risk' on their cybersecurity preparedness framework.

source

GPT 5.2 Pro (manually tested in ChatGPT by Epoch AI) gets 31% on FrontierMath Tier 4, beating the previous record of 19% by Gemini 3 Pro

source

Alibaba released Qwen3-Max-Thinking, their new flagship reasoning model.

They claim it directly competes with GPT-5.2 Thinking, Claude Opus 4.5 and Gemini 3 Pro.

https://qwen.ai/blog?id=qwen3-max-thinking

Sample document from Prism

Google finally released a public version of their Genie 3 world model.

https://labs.google/projectgenie

OpenAI to retire GPT-4o/4.1/4.1 mini/5 and o4-mini in ChatGPT. API access is unaffected.

https://openai.com/index/retiring-gpt-4o-and-older-models/

https://static.stepfun.com/blog/step-3.5-flash/

Step 3.5 Flash - new LLM by StepFun. 196B total, 11B active params, claims to be better at agentic programming than GLM 4.7 (which is 2x the size).

OpenAI is launching a standalone Codex app for interacting and building with agents.

It seems to be a direct answer to app versions of Claude Code/Cowork by Anthropic.

Available on macOS.

https://openai.com/codex/